AI Frontier Review

AI agents force governance to pivot from reviewing model outputs to tracking what systems actually do, why they were permitted, and who remains...

AI already delivers concrete oncology advances through pattern recognition and design, yet full discovery still requires human-led validation.

-...

Jais 3's six-model fleet plus speech recognition captures Arabic as actually used—dialects from Gulf to North Africa, code-switching with English or...

Anthropic's draft bill sets a high bar for government action: intervention only on foreseeable risks exceeding $1B in damages or 50+ serious injuries...

Employee-like access lets METR inspect training states, internal tools, and infrastructure at Anthropic, revealing whether training suppresses...

rStation lets agents from different users collaborate on problems like Millennium Prizes through shared forums and papers, yet explicitly leaves all...

• Secure institutional memory: Governance tools must integrate AI with the board's own record rather than public LLMs that lack history
  -...

Insurers determine AI autonomy by mapping claims tasks to risk tiers, requiring human sign-off on coverage and payment decisions while allowing AI to...

ASPIRE V1.0 enables real robot skill discovery and self-improvement, yet benchmarks jumped to ~99% overnight after switching the backend to Astra....

• Emergency planning gains: Improved 2-6 week forecasts give farmers, energy operators and emergency services more time to prepare for extremes like...

NVIDIA expanded its open Nemotron family to cover always-on agents and local hardware:

• Nemotron 3.5 Super VL (120B): multimodal agentic model for...

Operational metrics and structured frameworks are converging with advanced research methods to assess new LLMs reliably under data constraints.

-...

Krea 3 reveals what scalable editing workflows actually require from image models:

• Native regional edits with pixel-wise precision and minimal...

AI tools now produce superficially plausible papers that fabricate analyses of open neuroimaging datasets, inventing nonexistent statistics and...

OmniCompass introduces a holistic benchmark for omni-modal video grounding and understanding, directly testing how models reason across video combined with multiple modalities.

• EdgeWise runs a local SLM for most steps and delegates only critical ones to a cloud LLM, preserving performance while slashing inference costs.
  -...

Is AI interpretability close to a breakthrough?

Cammarata's essay argues neural networks are far cleaner than assumed, with most neurons having one...

Congressional scrutiny of the FAA’s new AI-enabled systems FMDS and SMART highlights how operational AI adoption in safety-critical aviation...

Capy-flash shows a 125B-parameter MoE model running on two 16 GB consumer GPUs, underscoring the potential to bring massive mixture-of-experts architectures to everyday hardware without specialized infrastructure.

CNN's multi-year deal grants Google enhanced content rights for generative AI experiments that surface its journalism to new audiences, with explicit...